Article starts from energy based approaches and comes to phase transition
Energy based defines probability as
\[p(E) = \frac{exp()}{Z}\]$Z$ - partition function .
Best papers in field comes from Yann LeCun
Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture paper
![]() |
---|
No we can’t handle. We need one more neuron |
Symmetry holds
For rotation in nD dimension we at least need n-1 parameters.
![]() |
---|
No we can’t handle. We need one more neuron |
Size of neural ensamble
Tightly connected with critical exponents, which are defined as
\[\]But it might be more intuitive from view of chemistry. Suppose we perfectly know everything about molecule. Every angle and atom of it’s structure
![]() |
---|
But what will be if we put them all together |
Main intuition is that radical changes of matter is connected with change of it’s energetic
\[H(X)\]$\Lambda$ is known is order parameter
Coherency length
\[\xi\]Brought from awesome video Renormalization: The Art of Erasing Infinity
Let’s solve:
\[\varepsilon x^2 + 2 x + 1 = 0\]Usin
A phase transition between positional and semantic learning in a solvable model of dot-product attention https://arxiv.org/abs/2402.03902
Article advices solution of problem and shows phase transition